[llvm-branch-commits] [llvm] 5ead757 - [VE] Support pack_f32p and pack_f32a intrinsic instructions

[Kazushi Marukawa via llvm-branch-commits]

Author: Kazushi (Jam) Marukawa
Date: 2021-01-08T22:59:11+09:00
New Revision: 5ead757f1d89898941f0e10f71f382aaf36d174b

URL: https://github.com/llvm/llvm-project/commit/5ead757f1d89898941f0e10f71f382aaf36d174b
DIFF: https://github.com/llvm/llvm-project/commit/5ead757f1d89898941f0e10f71f382aaf36d174b.diff

LOG: [VE] Support pack_f32p and pack_f32a intrinsic instructions

Support pack_f32p and pack_f32a intrinsic instructions and regression tests.

Reviewed By: simoll

Differential Revision: https://reviews.llvm.org/D94296

Added: 
    llvm/test/CodeGen/VE/VELIntrinsics/pack.ll

Modified: 
    llvm/include/llvm/IR/IntrinsicsVE.td
    llvm/lib/Target/VE/VEInstrIntrinsicVL.td

Removed: 
    


################################################################################
diff  --git a/llvm/include/llvm/IR/IntrinsicsVE.td b/llvm/include/llvm/IR/IntrinsicsVE.td
index 3a8eec6d4dd5..4e682eee7cc2 100644
--- a/llvm/include/llvm/IR/IntrinsicsVE.td
+++ b/llvm/include/llvm/IR/IntrinsicsVE.td
@@ -4,6 +4,13 @@
 let TargetPrefix = "ve" in {
   def int_ve_vl_svob : GCCBuiltin<"__builtin_ve_vl_svob">,
                        Intrinsic<[], [], [IntrHasSideEffects]>;
+
+  def int_ve_vl_pack_f32p : GCCBuiltin<"__builtin_ve_vl_pack_f32p">,
+                            Intrinsic<[llvm_i64_ty], [llvm_ptr_ty, llvm_ptr_ty],
+                                      [IntrReadMem]>;
+  def int_ve_vl_pack_f32a : GCCBuiltin<"__builtin_ve_vl_pack_f32a">,
+                            Intrinsic<[llvm_i64_ty], [llvm_ptr_ty],
+                                      [IntrReadMem]>;
 }
 
 // Define intrinsics automatically generated

diff  --git a/llvm/lib/Target/VE/VEInstrIntrinsicVL.td b/llvm/lib/Target/VE/VEInstrIntrinsicVL.td
index 6cec6fece8d8..3525484af108 100644
--- a/llvm/lib/Target/VE/VEInstrIntrinsicVL.td
+++ b/llvm/lib/Target/VE/VEInstrIntrinsicVL.td
@@ -5,6 +5,18 @@
 // SVOB pattern.
 def : Pat<(int_ve_vl_svob), (SVOB)>;
 
+// Pack patterns.
+def : Pat<(i64 (int_ve_vl_pack_f32p ADDRrii:$addr0, ADDRrii:$addr1)),
+          (ORrr (f2l (LDUrii MEMrii:$addr0)),
+                (i2l (LDLZXrii MEMrii:$addr1)))>;
+
+def : Pat<(i64 (int_ve_vl_pack_f32a ADDRrii:$addr)),
+          (MULULrr
+            (i2l (LDLZXrii MEMrii:$addr)),
+            (LEASLrii (ANDrm (LEAzii 0, 0, (LO32 (i64 0x0000000100000001))),
+                             !add(32, 64)), 0,
+                      (HI32 (i64 0x0000000100000001))))>;
+
 // LSV patterns.
 def : Pat<(int_ve_vl_lsv_vvss v256f64:$pt, i32:$sy, i64:$sz),
           (LSVrr_v (i2l i32:$sy), i64:$sz, v256f64:$pt)>;

diff  --git a/llvm/test/CodeGen/VE/VELIntrinsics/pack.ll b/llvm/test/CodeGen/VE/VELIntrinsics/pack.ll
new file mode 100644
index 000000000000..7ad702ace430
--- /dev/null
+++ b/llvm/test/CodeGen/VE/VELIntrinsics/pack.ll
@@ -0,0 +1,41 @@
+; RUN: llc < %s -mtriple=ve -mattr=+vpu | FileCheck %s
+
+;;; Test pack intrinsic instructions
+;;;
+;;; Note:
+;;;   We test pack_f32p and pack_f32a pseudo instruction.
+
+; Function Attrs: nounwind readonly
+define fastcc i64 @pack_f32p(float* readonly %0, float* readonly %1) {
+; CHECK-LABEL: pack_f32p:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ldu %s0, (, %s0)
+; CHECK-NEXT:    ldl.zx %s1, (, %s1)
+; CHECK-NEXT:    or %s0, %s0, %s1
+; CHECK-NEXT:    b.l.t (, %s10)
+  %3 = bitcast float* %0 to i8*
+  %4 = bitcast float* %1 to i8*
+  %5 = tail call i64 @llvm.ve.vl.pack.f32p(i8* %3, i8* %4)
+  ret i64 %5
+}
+
+; Function Attrs: nounwind readonly
+declare i64 @llvm.ve.vl.pack.f32p(i8*, i8*)
+
+; Function Attrs: nounwind readonly
+define fastcc i64 @pack_f32a(float* readonly %0) {
+; CHECK-LABEL: pack_f32a:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ldl.zx %s0, (, %s0)
+; CHECK-NEXT:    lea %s1, 1
+; CHECK-NEXT:    and %s1, %s1, (32)0
+; CHECK-NEXT:    lea.sl %s1, 1(, %s1)
+; CHECK-NEXT:    mulu.l %s0, %s0, %s1
+; CHECK-NEXT:    b.l.t (, %s10)
+  %2 = bitcast float* %0 to i8*
+  %3 = tail call i64 @llvm.ve.vl.pack.f32a(i8* %2)
+  ret i64 %3
+}
+
+; Function Attrs: nounwind readonly
+declare i64 @llvm.ve.vl.pack.f32a(i8*)